The Agentic Infrastructure Gap: In-Distribution Languages Make It a Coding Problem
summary
Claim: infrastructure fails because cloud consoles, API keys and copy‑pasted YAML are out of distribution for LLMs, so the fix is to express infra as real code they already understand. Evidence: SWE‑bench jumps from 33% to 86% in 21 months and Anthropic reaches 94%, Pulumi helps a frontier lab scale training/inference ~10x, Wiz manages >1M resources and >100k deployments/day, Composable AI runs 100% agentic infrastructure, and ~20% of customer deployments are already agentic. Practical pieces: Python infra for CoreWeave + NVIDIA GPU clusters, plan/preview diffs and typed modules to provide guardrails.
AI-written summary · about
chapters (15)
- 0:00Agentic coding in 2026
- 1:28Infrastructure bottleneck for AI agents
- 2:52Infrastructure management evolution
- 4:14From vibe coding to vibe engineering
- 6:00Coding thrives in-distribution
- 7:25Infrastructure gap and OOD challenges
- 9:53Infrastructure as a coding problem
- 11:19Infrastructure in code example
- 13:49Guardrails for agentic changes
- 15:35Infrastructure coding primitives
- 17:00Code scales with complexity
- 18:47AI-native dynamic workloads
- 19:48Agentic infrastructure examples
- 20:52Autonomous infrastructure path
- 21:56Risk and reward balance
related talks
12mThe State of Infrastructure-From-Code 2023
30mInfrastructure from Code: What Happened and What's Next
44mInfrastructure as Software
32mInfrastructure From Code: Next up for DevOps?
36mAI's Disruption of Infrastructure Operations
33mScaling Infrastructure as Code
44mInfrastructure from Code: The Next Generation of Cloud Management
33mHumans on the loop, not in it: Taking agentic engineering end to end